My solutions

.gitignore

@@ -0,0 +1,20 @@
+.env
+venv/
+.venv/
+__pycache__/
+
+
+# Ignore all GGUF models
+*.gguf
+
+# Ignore all FAISS index files
+*.faiss
+
+# Ignore chapter10 trained model outputs
+chapter10/augmented_embedding_model/
+chapter10/base_embedding_model/
+chapter10/cosineloss_embedding_model/
+chapter10/finetuned_embedding_model/
+chapter10/mnrloss_embedding_model/
+chapter10/results/
+chapter10/tsdae_embedding_model/

chapter06/Chapter 6 - Prompt Engineering.ipynb

@@ -1277,7 +1277,7 @@
    "provenance": []
   },
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": "hands-on-llms",
    "language": "python",
    "name": "python3"
   },
@@ -1291,7 +1291,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.12.9"
+   "version": "3.10.19"
   }
  },
  "nbformat": 4,

chapter06/chap6_notes.md

@@ -0,0 +1,120 @@
+# Chapter 6 - Prompt Engineering
+My notes.
+
+
+### Temperature
+Controls the randonmness or creativity.
+- 0: gerenates the same response, choose the most likely word
+- lower (0.2) for more deterministic output
+- higher (0.8) more diverse and creative
+- introduce stochastic behavior
+
+
+### top_p - nucleus sampling
+- sampling technique that controls wich subset of tokens (nucleus) the LLM can consider
+- top_p = 0.1, consider tokens until reaches that value
+- top_p = 1, consider all tokens
+
+
+We can combine each of those parameters for control the output of our LLM. 
+
+Use cases:
+- brainstorming: high temperature, high top_p
+- email-generation: low temperatura, low top_p
+- creative writing: high temperature, low top_p
+- translation: low temperature, high top_p
+
+## Prompts
+Basically are:
+- questions
+- statements or
+- instructions
+
+Basic prompt: 
+Input > LLM > Output
+
+### Instruction-based prompting
+- have the llm answer a specific question or resolve a certain task
+
+![alt text](images/image.png)
+
+
+Components of complex prompting:
+- persona: the role
+- instruction: the task
+- context: - additional information about the problem or task
+- format: the output format
+- audience: the target of the generated text
+- tone: the tone of voice
+- data: the data related to the task
+
+You can try many itterations over modular components.
+
+You can be as creative as possible:
+- use emotional stimuli: "this is very important for my job"
+- different models for different purposes
+
+
+**Zero-shot prompt**
+
+Prompting without examples
+
+    Classify the text into neutral, negative, or positive.
+
+    Text: The service was really slow. Sentiment: ...
+
+**One-shot prompt**
+
+Prompting with a single example
+
+    Classify the text into neutral, negative, or positive.
+
+    Text: The movie was fine. Sentiment: Neutral.
+
+    Text: The performance was outstanding. Sentiment:
+
+**Few-shot prompt**
+
+Prompting with more than one example
+
+    Classify the text into neutral, negative, or positive.
+
+    Text: The product broke after one day. Sentiment: Negative.
+
+    Text: This is the best purchase I've ever made! Sentiment: Positive.
+
+    Text: The shirt fits as expected. Sentiment: Neutral.
+
+    Text: The laptop is very fast, but the battery life is short. Sentiment:
+
+
+### Chaining prompts
+- allows llm to spend more time on individual question instead of tackling the whole problem
+- more calls to the model but we can give different parameters for each one
+- we can use the generated output as input
+
+## Reasoning
+
+![alt text](images/image2.png)
+
+### Chain-of-Thought
+- think before answering
+- 'let's think step-by-step'
+
+### Self-consistency
+- sampling outputs
+- multiple reasoning paths
+- ask the same prompt multiple times and take the majority result as final answer
+
+![alt text](images/image3.png)
+
+### Tree-of-Thought
+- tree-based structure
+- generate intemediate thoughts to be rated
+- the most promising are kept and the lowest are pruned
+- we can emulate a conversation between multiple experts
+
+
+## Evaluate outputs
+- expected format
+- if the output follows out rules